Interesting idea! Perhaps the 'random' docs could instead be generated (or even trained) for even-greater significance of the new embeddings. For example: aft

Interesting idea! Perhaps the 'random' docs could instead be generated (or even trained) for even-greater significance of the new embeddings. For example: after doing LDA, generate a 'paragon' doc of each topic. Or coalescing all docs of a known label together, then reducing them to D summary pseudo-words – the D 'words' with minimum total WMD to all docs of the same label. Or adding further R docs into regions of maximum confusion.